Avoiding EBUSY on TPM writes

We're running into a situation where concurrent TPM accesses can trigger EBUSY on write due to there either being a queued command or a response not yet having been read. Obviously one approach would be to retry, but that involves us either spinning or waiting for an arbitrary amount of time between attempts, which doesn't seem ideal. There's a poll interface that tells us whether there's a response to read but (a) that doesn't help in the enqueued command situation and (b) this would still be racy - we could be notified that the response has been read, be preempted, and have another process perform a write before we get the opportunity to.

What's the right way to fix this? One approach would simply be to replace the EBUSY with an interruptible sleep and wake the process when the TPM is available, but that feels like it's technically an ABI break.

That's by design: the fd interface is command/response per open fd, so if you try to send another command before you get a response you'll get an EBUSY by design. If you didn't, how would you know which response belonged to which command (there's no counter or other identifier)?

However, this behaviour is per-fd, why can't you just open multiple fd's for your multiple accesses? /dev/tpmrm0 will assure that they don't see each other and that the commands are properly sequenced without giving you an EBUSY. And bonus you don't have to keep global track of how many transient resources you've used.

James

Argh! Sorry, I'd missed that this was per-fd - that makes complete sense. Let me figure out why we're re-using the same fd for multiple requests.